Skip to content

README-entry-gpu-sizing.md

Latest commit

 

History

History
26 lines (18 loc) · 2.88 KB

README-entry-gpu-sizing.md

File metadata and controls

26 lines (18 loc) · 2.88 KB
Raw

GPU sizing card — Life Sciences / Chemistry / Materials Science agents

Purpose. Right-sized GPU recommendations for each agent: the cheapest GPU class and memory size that does not bottleneck a general-purpose run. Intended as a practical replacement for the registry's recommended_sku field, which is defined in tool-definition-schema.json as an ordered preference ladder rather than a hardware requirement, and is frequently read as the latter by customers and partners.

Scope. 41 of the 46 agents in agent-registry.json — the life sciences and chemistry / materials science set. The silicon (Yosys, OpenSTA, Icarus Verilog, Xyce) and CFD (OpenFOAM) agents are excluded; all five are CPU-only and none has a GPU SKU. Of the 41: 6 require a GPU, 8 benefit from one, 24 are CPU-only, and 3 are LLM-only with no tool container.

Assumptions.

  • Sizing targets single-job, general-purpose inference — not model training, and not campaign-scale batch throughput, both of which justify larger cards.
  • fp16 or bf16 precision where the tool supports it. Note that Turing (T4) has fp16 tensor cores but no bf16.
  • CPU and host-RAM floors are taken from each tool's existing min_resources and were not independently re-derived.

Limitations.

  • Not benchmarked in-house across every GPU class. Recommendations rest on published benchmarks where they exist and on model architecture where they do not. Corrections from hands-on experience are welcome and expected.
  • Three sizings — BoltzGen, RetroChimera, TamGen — are reasoned from architecture rather than measured, and are marked (est.) in the card.
  • Snapshot of registry commit d2ff532 (generated 2026-06-01). Compute blocks are hand-authored per agent and change independently, so re-check against main before circulating.
  • Two hard floors are not negotiable by price: six agents set min_resources.gpu = 1 and will not start without a GPU, and Boltz-2 and BoltzGen require Ampere or newer (the trifast kernels do not run on V100-class hardware).
  • recommended_sku is advisory metadata — the supercomputer CLI does not read it and nothing validates it against the workload. Changing it is a pre-deployment edit to the tool's tool.yaml, not a platform setting.

Sources.

  • .auto-registry/agent-registry.json — agent list, associated tools
  • agents/*/tools/*/tool.yaml — compute.min_resources, max_resources, recommended_sku, pool_type
  • docs/schemas/tool-definition-schema.json — definition of recommended_sku as an ordered list
  • Primary literature, cited per row in the card. Benchmarks with explicit GPU hardware include Passaro et al. (bioRxiv 2025, Boltz-2, H100), Zeni et al. (Nature 2025, MatterGen, 8× A100 training), Carnimeo et al. (JCTC 2023, Quantum ESPRESSO, A100 80 GB), Páll et al. (JCP 2020, GROMACS, V100 PCIe), and Solis-Vasquez et al. (Parallel Computing 2021, AutoDock-GPU, A100/V100/RTX2070).